The MSM website provides a program package that allows nutritional scientists to calculate usual dietary intakes by combining short-term and long-term measurements (multiple sources). It promotes simple access to the MSM to estimate usual food intake for individuals and populations.

Estimating usual food intake distributions from short-term quantitative measurements is critical when occasionally or rarely eaten food groups are considered. To overcome this challenge by statistical modeling, the Multiple Source Method (MSM) was developed in 2006. The MSM provides usual food intake distributions from individual short-term estimates by combining the probability and the amount of consumption with incorporation of covariates into the modeling part. Habitual consumption frequency information may be used in 2 ways: first, to distinguish true nonconsumers from occasional nonconsumers in short-term measurements and second, as a covariate in the statistical model. The MSM is therefore able to calculate estimates for occasional nonconsumers. External information on the proportion of nonconsumers of a food can also be handled by the MSM. As a proof-of-concept, we applied the MSM to a data set from the European Prospective Investigation into Cancer and Nutrition (EPIC)-Potsdam Calibration Study (2004) comprising 393 participants who completed two 24-h dietary recalls and one FFQ. Usual intake distributions were estimated for 38 food groups with a proportion of nonconsumers > 70% in the 24-h dietary recalls. The intake estimates derived by the MSM corresponded with the observed values such as the group mean. This study shows that the MSM is a useful and applicable statistical technique to estimate usual food intake distributions, if at least 2 repeated measurements per participant are available, even for food groups with a sizeable percentage of nonconsumers.

Background/Objectives: The aim of this paper was to compare methods to estimate usual intake distributions of nutrients and foods. As 'true' usual intake distributions are not known in practice, the comparison was carried out through a simulation study, as well as empirically, by application to data from the European Food Consumption Validation (EFCOVAL) Study in which two 24-h dietary recalls (24-HDRs) and food frequency data were collected. The methods being compared were the Iowa State University Method (ISU), National Cancer Institute Method (NCI), Multiple Source Method (MSM) and Statistical Program for Age-adjusted Dietary Assessment (SPADE). Subjects/Methods: Simulation data were constructed with varying numbers of subjects (n), different values for the Box-Cox transformation parameter (l BC ) and different values for the ratio of the within-and between-person variance (r var ). All data were analyzed with the four different methods and the estimated usual mean intake and selected percentiles were obtained. Moreover, the 2-day within-person mean was estimated as an additional 'method'. These five methods were compared in terms of the mean bias, which was calculated as the mean of the differences between the estimated value and the known true value. The application of data from the EFCOVAL Project included calculations of nutrients (that is, protein, potassium, protein density) and foods (that is, vegetables, fruit and fish). Results: Overall, the mean bias of the ISU, NCI, MSM and SPADE Methods was small. However, for all methods, the mean bias and the variation of the bias increased with smaller sample size, higher variance ratios and with more pronounced departures from normality. Serious mean bias (especially in the 95th percentile) was seen using the NCI Method when r var ¼ 9, l BC ¼ 0 and n ¼ 1000. The ISU Method and MSM showed a somewhat higher s.d. of the bias compared with NCI and SPADE Methods, indicating a larger method uncertainty. Furthermore, whereas the ISU, NCI and SPADE Methods produced unimodal density functions by definition, MSM produced distributions with 'peaks', when sample size was small, because of the fact that the population's usual intake distribution was based on estimated individual usual intakes. The application to the EFCOVAL data showed that all estimates of the percentiles and mean were within 5% of each other for the three nutrients analyzed. For vegetables, fruit and fish, the differences were larger than that for nutrients, but overall the sample mean was estimated reasonably. Conclusions: The four methods that were compared seem to provide good estimates of the usual intake distribution of nutrients. Nevertheless, care needs to be taken when a nutrient has a high within-person variation or has a highly skewed distribution, and when the sample size is small. As the methods offer different features, practical reasons may exist to prefer one method over the other.

Assessment of human cancer risk from animal carcinogen studies is severely limited by inadequate experimental data at environmentally relevant exposures, and procedures requiring modeled extrapolations many orders of magnitude below observable data. We used rainbow trout, an animal model well suited to ultra low-dose carcinogenesis research, to explore dose-response down to a targeted 10 excess liver tumors per 10,000 animals (ED001). A total of 40,800 trout were fed 0–225 ppm dibenzo[a,l]pyrene (DBP) for four weeks, sampled for biomarker analyses, and returned to control diet for nine months prior to gross and histologic examination. Suspect tumors were confirmed by pathology, and resulting incidences were modeled and compared to the default EPA LED10 linear extrapolation method. The study provided observed incidence data down to two above-background liver tumors per 10,000 animals at lowest dose (that is, an un-modeled ED0002 measurement). Among nine statistical models explored, three were determined to fit the liver data well - linear probit, quadratic logit, and Ryzin-Rai. None of these fitted models is compatible with the LED10 default assumption, and all fell increasingly below the default extrapolation with decreasing DBP dose. Low-dose tumor response was also not predictable from hepatic DBP-DNA adduct biomarkers, which accumulated as a power function of dose (adducts = 100 * DBP1.31). Two-order extrapolations below the modeled tumor data predicted DBP doses producing one excess cancer per million individuals (ED10−6) that were 500–1500-fold higher than that predicted by the five-order LED10 extrapolation. These results are considered specific to the animal model, carcinogen, and protocol used. They provide the first experimental estimation in any model of the degree of conservatism that may exist for the EPA default linear assumption for a genotoxic carcinogen.